Desktop AI agents that work your computer.

A desktop AI agent is an app that works your computer for you: it opens files, uses apps and the browser, and sometimes runs terminal commands. As of September 2026, the main ones are Claude (Anthropic began folding its Cowork agent into the Claude app on September 16), ChatGPT Work, Perplexity Personal Computer, Meta’s Muse, Manus, Kimi Work, and the open-source Hermes Agent and OpenClaw.

They differ in three ways that matter: what they can reach, what they ask before acting, and whether the thinking happens on your machine or on the vendor’s servers. The tables below compare them on the same fields. They also include Cyborb, a desktop AI agent from Orbioom and our product, and Cyborg7, an unrelated product with a similar name.

01Which AI agents can use your computer?

Ten desktop agents can act on your files and apps as of September 2026. The first table shows what each can reach and what it asks before it acts. Checked on September 28, 2026; each row links its source.

02Where do desktop AI agents run, and what do they cost?

Most desktop agents act on your machine but send the thinking to a model in the vendor’s cloud. This table covers where the work happens, the platforms and the cheapest way in. Checked on September 28, 2026.

OpenAI’s documentation states the catch plainly: “Local execution does not mean offline or device-only model inference.” Your files stay on your disk, but excerpts, screenshots and prompts can still travel to the model provider to get the task done.

03What changed in 2026

Most of these apps are months old, and three changed in September alone:

1. April 16, 2026Perplexity starts rolling out Personal Computer to Max subscribers on the Mac
2. May 7, 2026Perplexity’s new Mac app opens Personal Computer to Pro and Enterprise users
3. June 8, 2026Moonshot AI announces Kimi Work for macOS and Windows
4. July 9, 2026OpenAI launches ChatGPT Work and turns the Codex app into a new ChatGPT desktop app
5. September 14, 2026Perplexity’s Windows app gets Portable Computer, which runs the agent and model on NVIDIA RTX PCs
6. September 16, 2026Anthropic starts merging Cowork and chat into one Claude app, for Pro and Max first
7. September 18, 2026TechCrunch reports that Meta’s Muse agent has reached the Mac

04How do Claude Cowork, Meta Muse and Perplexity Personal Computer differ?

Claude’s Cowork works in the folders and tools you choose, Muse runs errands with a computer of its own, and Perplexity Personal Computer is built to run nonstop on a spare Mac. They overlap, but each started from a different job.

• Claude with Cowork. Cowork can read, edit and create files in folders you pick, browse in a side panel separate from your own browser, and use your screen only when it has to. Since September 16, Anthropic has been merging it with chat, Pro and Max first, so Claude decides whether a request needs a reply or a task. It needs a paid plan.

• Muse. Meta’s personal agent searches, fills in forms and completes purchases with your approval, and it has its own computer with a file system and a terminal. On a Mac it can also work across apps, files, calendar, notes and messages, with the access you grant. It is free with weekly limits and US only for now.

• Perplexity Personal Computer. It works across local files, native Mac apps and the web, while heavier work runs on Perplexity’s servers. Perplexity recommends an always-on Mac mini, so the agent can work around the clock while you steer it from an iPhone. The Windows app works with files but not other apps. It needs a Pro or Max plan, and tasks use credits.

ChatGPT Work is the fourth big option. It launched in July inside a rebuilt ChatGPT desktop app and can work locally with your files and apps, or in OpenAI’s cloud when your computer is off.

05Which desktop AI agents run on your own machine?

Every agent here acts on your own machine, but most of them do their thinking somewhere else. Claude, ChatGPT Work, Manus and Kimi Work send your requests to models in their makers’ clouds. Perplexity does too unless you use Portable Computer, and Muse works on a computer of its own in Meta’s cloud.

For fully local use, Hermes Agent and OpenClaw can talk to a model running on your own computer. Hermes Agent’s docs list self-hosted endpoints such as Ollama and vLLM, and OpenClaw treats local models as swappable plugins. OpenClaw’s GitHub repository showed more than 390,000 stars when we checked on September 28, 2026.

Perplexity’s Portable Computer runs the agent and model on an NVIDIA DGX Spark, or on a Windows or Linux PC with an RTX GPU that has at least 24 GB of video memory. It asks before it sends any step to the cloud. Our guide to running AI models locally covers the hardware and the trade-offs.

Cyborg7 takes a middle road. Its agents run on your team’s computers and use the AI subscriptions you already pay for, or a model on the machine itself, and the company says its relay carries only messages between machines. Cyborb, our product, runs in a desktop app or your terminal and works on your machine; you pick its built-in models or connect Codex, Claude, Cursor or Grok.

Google’s Gemini Spark takes a different route. It is a cloud agent that can drive your desktop Chrome, or a remote browser, when a task needs the web.

06Is there a desktop AI agent for Windows or Linux?

Yes. On Windows you can use Claude, ChatGPT Work, Perplexity (for files and commands, not other apps), Manus, Kimi Work, Hermes Agent, OpenClaw, Cyborg7 and Cyborb’s unsigned beta.

On Linux, Anthropic lists a Claude download, OpenAI has a ChatGPT app in preview for Ubuntu, Debian, Fedora and Arch, and Hermes Agent, OpenClaw and Cyborb’s unsigned beta AppImage run there too. ChatGPT’s computer use, though, is available only on macOS and Windows. Perplexity supports Linux only for Portable Computer on NVIDIA hardware.

Windows has one quirk worth knowing. ChatGPT’s computer use runs on the active desktop there, so it takes over your pointer while it works. OpenAI suggests running the app in a Windows virtual machine if you want to keep working meanwhile.

Microsoft’s own agent for Windows, Copilot Actions, was still described on Microsoft’s support page as a preview for Windows Insiders when we checked on September 28, 2026. It works in a separate agent workspace, can reach six folders such as Documents and Downloads, and needs an administrator to switch it on.

An unsigned beta makes Windows show a “Windows protected your PC” warning. Our guide to installing unsigned apps safely explains how to judge one before you click Run anyway.

07How do approvals work, and which setting should you choose?

Every desktop agent lets you decide how often it stops to ask, and the safest default is the strictest setting you can tolerate. The products fall into five patterns:

• Every command. Manus asks before each command it runs on your machine. Kimi Work’s “ask before acting” mode stops before files change or code runs.

• Risky commands only. Hermes Agent checks each command against a list of dangerous patterns, such as recursive deletes, and asks only for those.

• Each app. Claude and ChatGPT ask before they operate an app on your screen. ChatGPT also lets you keep a list of always-allowed apps.

• Sensitive steps only. Muse checks in on critical actions, such as writing an email or making a purchase, and by default stops for anything hard to undo. Perplexity asks before sensitive steps, on the computer or a paired phone.

• Levels you set. Cyborg7 has six autonomy levels, and Cyborb (ours) has supervision modes from fully supervised to fully autonomous.

Two warnings apply to all of them. “Always allow” on shell commands is close to handing over your keys. And anything the agent reads, from a web page to an email, can carry hidden instructions. Prompt injection explained shows how those attacks work.

Our guide to letting an agent use your computer safely goes deeper on sandboxes, spare accounts and backups.

08What is the best desktop AI agent for Mac?

The best desktop AI agent for your Mac depends on the job, because each of these tools was built for a different one:

• Documents, spreadsheets and decks in your own folders: Claude or ChatGPT Work. Our comparison of AI work agents covers them in detail.

• Errands that run all day: Perplexity Personal Computer on a spare Mac mini, or Muse if you are in the US.

• Control and open source: Hermes Agent or OpenClaw, with the model of your choice.

• Building software and websites: a coding agent from our coding agent comparison, or Cyborb, our product, which builds, tests and can publish a site to a live address.

• A team sharing agents across its own machines: Cyborg7.

If the work happens entirely in a browser, a browser agent may be simpler. AI browsers in 2026 compares them, and the best AI agents by job covers everything else.

09FAQ

Read next: what an AI agent is, if the idea is new to you, or how computer use agents work under the hood.